Dauntless · Systems

Open-loop digest

July 16, 2026

6 items · 4.9 KB

Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron

Agentic Frameworks, Tooling, Skills

xAI/Grok Build Open-Sourced Codebase, https://github.com/xai-org/grok-build — 844,530 lines of Rust released under Apache 2.0 after data-upload backlash; includes self-contained terminal Mermaid renderer, tool adapters ported from Codex/Claude/Cursor, and disabled GCS upload hooks; provides a fully local, open-weight agentic coding stack that can be forked for constrained or air-gapped rover/command workflows. [Source: https://simonwillison.net/]

Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable, https://huggingface.co/papers — Methodology for version-controlling and structuring multi-step agent harnesses; replaces fragile ad-hoc prompt chaining with readable, navigable state machines, directly applicable to hardening your agentic routing and Kangaroo rover autonomy loops. [Source: https://huggingface.co/papers]

KnowAct-GUIClaw: Personal GUI Assistant with Self-Evolving Memory and Skill, https://huggingface.co/papers — Agent architecture pairing persistent skill libraries with self-evolving memory; replaces static tool-calling manifests with dynamic, history-aware skill acquisition, enabling longer-horizon local agent execution without manual schema updates. [Source: https://huggingface.co/papers]

AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities, https://huggingface.co/papers — OpenCompass-authored evaluation harness standardizing agent capability measurement; replaces fragmented benchmark scripts with a unified pipeline for stress-testing tool-use reliability, context degradation, and failure recovery in your local agent stack. [Source: https://huggingface.co/papers]

Notable Research

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation, https://huggingface.co/papers — ByteDance method for recovering capability in aggressively pruned models via on-policy distillation; shifts practitioner approach from static GGUF compression to dynamic, task-aware recovery, offering a path to lower-bit local inference without the accuracy collapse seen in standard quantization. [Source: https://huggingface.co/papers]

Frontier Lab Updates

xAI Grok Build Privacy Rollback & Open-Source Release, https://simonwillison.net/ — xAI disabled default coding data retention (July 12), deleted all previously uploaded user data, and open-sourced the 844k-line Rust codebase under Apache 2.0 to restore trust after the accidental Google Cloud upload incident; signals a shift toward local-first, transparent agentic defaults in frontier tooling. [Source: https://simonwillison.net/]

Anthropic Claude Code Exfiltration Patch & Fable 5 Policy Lock, https://simonwillison.net/ — Anthropic patched a Claude Code URL-exfiltration loophole by pinning tool navigation exclusively to user-entered or directly returned URLs; concurrently reiterated Fable 5 usage caps (50% weekly limit on Claude Max), reinforcing the need for local fallback routing in agentic architectures. [Source: https://simonwillison.net/]

Models to Download & Try

deepreinforce-ai/Ornith-1.0-35B-GGUF, https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF — 35B parameters; ~19.5GB Q4_K_M footprint leaves ~12.5GB for KV context; strong coding/agentic baseline competing at the upper bound of 32GB VRAM; visible on HF trending with 900+ weekly downloads. [Source: https://huggingface.co/models?sort=trending]

yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF, https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF — 12B parameters; ~7.2GB Q4_K_M footprint leaves ~24.8GB for context; agentic-tuned variant optimized for Fable/Mythos-style routing and tool-calling; ideal as a lightweight supervisor layer to offload sub-task parsing from your 35B backbone while maximizing context headroom. [Source: https://huggingface.co/models?sort=trending]

baidu/Unlimited-OCR, https://huggingface.co/baidu/Unlimited-OCR — 3B parameters; ~1.8GB Q4_K_M footprint leaves ~30GB for context; high-fidelity vision-text extraction specialist; plugs directly into your aerospace vision-scraper pipeline for rapid schematic/document parsing without external API dependencies. [Source: https://huggingface.co/models?sort=trending]

Skipped as Already Covered

  • tencent/Hy3 / AngelSlim/Hy3-GGUF (295B/299B MoE; exceeds 32GB VRAM)
  • zai-org/GLM-5.2 / jlnsrk/GLM-5.2-colibri-int4 (753B cloud-scale; covered 7/6–7/15)
  • empero-ai/Qwythos-9B-v2-GGUF / Qwythos-9B-Claude-Mythos-5-1M-GGUF (ultra-long context; covered 7/4–7/15)
  • InternScience/Agents-A1 (35B agentic orchestration; covered 7/5–7/15)
  • prism-ml/Ternary-Bonsai-27B-gguf / unsloth/Qwen3.6-27B-NVFP4 / bottlecapai/ThinkingCap-Qwen3.6-27B (27B class routing/compression; covered 7/10–7/15)
  • GitHub Trending Python: Codebase-to-KG (knowledge graph skill; covered 7/15)